Cherry Blossom Peak Bloom Prediction (Team 5103)

Setup

Competition context

• Predict next-year peak bloom day-of-year (DOY) for 5 sites.
• Evaluation: point accuracy (absolute error) + interval quality (coverage, then width).
• Full rules and framing are in README.md.

What this project does

• Reproducible R workflow in solution.qmd
• Independent Python check in Solution.ipynb
• Final outputs:
  • cherry-predictions.csv
  • cherry-predictions-python.csv
  • cherry-predictions-final.csv

Data used

• Core competition data: Kyoto, Washington DC, Liestal, Vancouver, NYC
• Auxiliary data: Japan regional, MeteoSwiss, South Korea, USA-NPN (NYC)
• Data structure and licensing notes: data/README.md

Load core data

# A tibble: 1,080 × 7
   source      location   lat  long   alt  year bloom_doy
   <chr>       <chr>    <dbl> <dbl> <dbl> <int>     <dbl>
 1 competition kyoto     35.0  136.    44   812        92
 2 competition kyoto     35.0  136.    44   815       105
 3 competition kyoto     35.0  136.    44   831        96
 4 competition kyoto     35.0  136.    44   851       108
 5 competition kyoto     35.0  136.    44   853       104
 6 competition kyoto     35.0  136.    44   864       100
 7 competition kyoto     35.0  136.    44   866       106
 8 competition kyoto     35.0  136.    44   869        95
 9 competition kyoto     35.0  136.    44   889       104
10 competition kyoto     35.0  136.    44   891       109
# ℹ 1,070 more rows

Exploratory data analysis: long-term trend

Data enrichment for NYC (USA-NPN)

# A tibble: 1 × 2
   year bloom_doy
  <dbl>     <dbl>
1    NA        98

Features used in modeling

• Time: year, centered year, squared year
• Geography: latitude, longitude, altitude (log-transformed)
• Data reliability proxy: site observation count
• Source indicator (competition vs auxiliary vs NPN)

(Implemented in add_features in solution.qmd.)

Model A: local trend

• Site-wise recency-weighted quadratic (fallback to linear/mean when sparse)
• Captures local momentum and curvature
• Implemented in predict_local_trend in solution.qmd

Model B: pooled nonlinear model

• R pipeline: GAM with smooths over year, spatial terms, altitude, site depth
• Implemented in fit_gam_model in solution.qmd
• Python pipeline: Gradient Boosting Regressor in Solution.ipynb

Backtesting and ensemble blending

• Rolling-origin backtest over historical years
• Compute MAE for local and pooled models
• Blend weights via inverse-MAE:
  • better out-of-sample model gets larger weight
• Implemented in rolling section of solution.qmd

Prediction intervals

• Split-conformal style calibration
• 90th percentile of absolute residuals by location
• Fallback to global residual quantile when needed

Final predictions (from file)

# A tibble: 5 × 4
  location     prediction lower upper
  <chr>             <dbl> <dbl> <dbl>
1 kyoto                90    80   100
2 liestal              88    78    96
3 newyorkcity          92    85   100
4 vancouver            92    77   108
5 washingtondc         83    76    90

Interpretation

• Ensemble improves robustness by combining:
  1. local historical behavior
  2. cross-site transferable structure
• Intervals are calibrated to target coverage while controlling width
• Approach is reproducible and competition-aligned

Limitations and next steps

• Residual weather shocks remain hard to predict
• Potential improvements:
  • engineered temperature covariates
  • richer uncertainty modeling
  • model stacking with strict out-of-sample validation

Reproducibility

• Render slides:

quarto render slides.qmd

• Main artifacts:
  • solution.qmd
  • Solution.ipynb
  • README.md
  • data/README.md